/Tech3h ago

Meta bans developers from using Claude Code and OpenAI Codex to prevent distillation into its own MetaCode

Story Overview

Meta is restricting employee access to Anthropic’s Claude Code and OpenAI’s Codex while building its internal coding model, MetaCode, to stop outputs from those tools from slipping into its training data or evaluations.

795983513462.2K

#201

Original post

Chubby♨️@kimmonismus#1360inTech

Meta is now facing the exact problem every AI company will soon face.

It wants to replace expensive external coding tools like Claude Code and Codex with its own internal system, MetaCode. But to build a better coding model, Meta has to make sure it is not accidentally training or evaluating on outputs from rival models.

That is the distillation trap: The more companies rely on frontier models to build internal AI infrastructure, the harder it becomes to prove where the intelligence actually came from.

6:48 AM · Jun 29, 2026 · 49.8K Views

Developer Impact

Engineers Navigate Tighter Rules

Applied AI teams at Meta must now avoid heavy reliance on the external tools they have been using to speed up work, creating friction during the transition to an in-house alternative.

Open Question

Distillation Risk Remains Unmeasured

Public details do not yet show how strictly the guidelines are enforced or whether they will actually keep rival model outputs from influencing MetaCode.

Sentiment

Positive users see potential upside in Meta's tricky balance restricting Claude and Codex to protect training data, while negative users call out the hypocrisy and absurd inconsistencies of the policy.

Pos

55.0%

Neg

45.0%

21 comments with sentiment.

Cluster Engagement

Digg Deeper

No Digg Deeper questions have been answered for this story yet.

THE INFORMATIONVia

#1360

Posts from X

Most Activity

VIEWS5.3KRETWEETS3

Chubby♨️@kimmonismus

"Meta, which has been among the biggest customers of Claude Code, set up the applied AI engineering team earlier this year and has tasked it with improving Meta’s own coding assistant, MetaCode. A big part of that entails building high-quality datasets and programming challenges that engineers use to train and test its models for coding work. While Meta permits the team to use outside AI tools for some purposes, it requires engineers to design those challenges themselves and to rely on their own technical expertise rather than AI-generated concepts."

Chubby♨️@kimmonismus

Meta is now facing the exact problem every AI company will soon face.

That is the distillation trap: The more companies rely on frontier models to build internal AI infrastructure, the harder it becomes to prove where the intelligence actually came from.

3h5.3K242

BOOKMARKS4LIKES33REPLIES5

Alexander Doria@Dorialexander

Still don't know why so many people take it as a settled matter while "distillation" is a much more fringe legal issue than training on copyright outputs. If fair use holds here, it holds everywhere.

Chubby♨️@kimmonismus

Meta is now facing the exact problem every AI company will soon face.

That is the distillation trap: The more companies rely on frontier models to build internal AI infrastructure, the harder it becomes to prove where the intelligence actually came from.

1h2.9K334

Chubby♨️@kimmonismus

Source: https://www.theinformation.com/articles/internal-docs-show-meta-putting-limits-claude-codex-fearing-distillation?rc=bfliih

Chubby♨️@kimmonismus

Meta is now facing the exact problem every AI company will soon face.

That is the distillation trap: The more companies rely on frontier models to build internal AI infrastructure, the harder it becomes to prove where the intelligence actually came from.

3h4.9K123

xlr8harder@xlr8harder

We should be more clear about this.

It might be time again for some stickers.

Alexander Doria@Dorialexander

Still don't know why so many people take it as a settled matter while "distillation" is a much more fringe legal issue than training on copyright outputs. If fair use holds here, it holds everywhere.

1h40190

Simon Willison@simonw

@Dorialexander This doesn't look to me like a fair use legal issue - it reads like a terms of service issue, where Meta have agreed to contractual terms with those third parties that say they can't use those models to train their own

Alexander Doria@Dorialexander

Still don't know why so many people take it as a settled matter while "distillation" is a much more fringe legal issue than training on copyright outputs. If fair use holds here, it holds everywhere.

32m43341

Alexander Doria@Dorialexander

@simonw I mean Meta intently trained on Libgen (and hosted it). Everything is a risk calculation.

Simon Willison@simonw

19m9930

Salio@Mr_Salio

@kimmonismus AI Regulation before GTA 6

3h1402

Nav Toor@heynavtoor

@kimmonismus AI companies now have to prove what they copied

41m124

Patrick Kuhnke@ku_ds17868

@kimmonismus wait metacode is a thing? just internally right?

3h111

Stumblinz@Stumblinz

@kimmonismus @grok any more well known news outlets echoing this?

2h71

Vincent Adultman@TatataToddC

@fvd_explore @kimmonismus I think it's mostly to avoid lawsuits

Anthropic had to pay billions because of a copyright lawsuit. If Meta distills Anthropic, they might suffer the same fate as they won't be able to prove where their training data came from.

2h111

Chubby♨️@kimmonismus

@Mr_Salio :(

3h1522

Mike Key@1337hero

@kimmonismus Wonder if we're going to see the photocopy problem with distillation.

How many times can you copy intelligence before the quality breaks down?

2h912

Chubby♨️@kimmonismus

@ku_ds17868 yes, i was surprised as well. Probably just internal model

3h1381

xlr8harder@xlr8harder

Obvious, right?

1h981

apolinario (poli)@multimodalart

@Dorialexander also, if folks train a model on "public github repos" nowadays, there will be so much claude/codex code there that avoid "distillation" is pretty hard

someone would need to do a "pangram for code", exclude all ai-code to then train a model

39m411

Tornado guy@fanofaliens

@kimmonismus They were literally using Gemini Until they blocked their usage

2h391

Priya@Priyannkaaaa

@kimmonismus

2h114

Dear AI 🥔@DearXson

@kimmonismus meta spent years copying features from other apps

now they can't even copy AI without getting caught

2h231

Adam Crabtree@CrabDude

@kimmonismus This is exactly why it will not remain defensible or matter long term. Lawsuits will happen, but they will all be decelerationist & will lose bc of social politics. Plea deals & negotiated bargains will prevail nearterm as the trolls take their tolls, but this too shall pass

1h69